Pick a pilot. Watch every decision.

An LLM flies a simulated flying car, one command every 0.2 s. On a running clock, the car keeps flying while the model thinks.

0.00 s

Command it sent

—

Telemetry it saw

All decisions

    About these flights

    Every flight comes from the test set: 192 flights per stunt in light wind and in gale, 384 in a storm, flown at temperature 0. Each wind level shows the first flight where the trained model, DeepSeek with unlimited time and the scripted pilot each do what they do on most flights. In light wind and gale the trained model was flown with the sim waiting; it answers in under 0.1 s, and its storm flights were identical with and without the clock. The scripted pilot is re-flown here.

    Replays are precomputed with the deterministic simulator; this page calls no model. Each telemetry line is rebuilt by replaying the pilot's logged commands and checked against the recorded flight.

    Fly your own model

    Run these flights with any model on SkyPilot Tokens. New accounts get $5 in free credits; during the preview, some are waitlisted.

    Get SkyPilot Tokens
    Show command 
    export SKYPILOT_TOKENS_API_KEY=...
python tools/pilot_bench.py --model deepseek-ai/DeepSeek-V4.1-Flash-1017 --realtime --per-kind 4 --out /tmp/my_pilot
    Run it in llm/flying-roadster-rl.

    Train your own model

    The 4B pilot here took about $10 of RL post-training on one H100. We can help you do the same for your own task.

    Talk to us